💻 computer science

Fourier Double Gate Physics-Informed Neural Networks for Multiscale PDE Solutions and Spatiotemporal Field Modeling

This paper proposes the Fourier Double Gate Physics-Informed Neural Network (FDG-PINN), which integrates multiscale Fourier feature encoding, a Double Gate architecture, and adaptive loss weighting to overcome spectral bias and gradient imbalance, thereby achieving superior accuracy in solving multiscale partial differential equations and reconstructing complex spatiotemporal temperature fields.

Original authors: Guangzheng Zhu, Changgui Gu, Hailing Wang

Published 2026-09-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guangzheng Zhu, Changgui Gu, Hailing Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict the weather or model how heat moves through a complex machine. Scientists often rely on mathematical rules called partial differential equations to describe these physical processes. For decades, solving these equations has required breaking the world down into a rigid grid, like a chessboard, and calculating the answer for every single square. This method works well for simple shapes but becomes incredibly difficult when the terrain is irregular or the data is sparse. In recent years, a new approach called Physics-Informed Neural Networks has emerged. Instead of a grid, these systems use artificial intelligence to learn the solution directly, guided by the laws of physics. However, these digital learners have a stubborn habit: they are naturally good at learning smooth, slow-changing patterns but struggle terribly with rapid, complex fluctuations. When a problem involves both slow trends and fast, jagged changes at the same time, the standard AI models often get stuck, failing to capture the full picture.

A team of researchers from the University of Shanghai for Science and Technology and Yunnan Normal University has developed a new method to fix this specific weakness. They created a system they call the Fourier Double Gate Physics-Informed Neural Network. The name sounds technical, but the idea is straightforward. The researchers realized that to teach an AI to see both the big picture and the fine details, they needed to change how the computer "sees" the input data. They added a special layer that translates the raw coordinates of space and time into a mix of different frequencies, much like how a prism splits white light into a rainbow of colors. This allows the network to immediately recognize both the slow, rolling waves of a temperature field and the sharp, quick spikes that usually confuse standard models.

To handle this new, richer information, the team also redesigned the internal structure of the network. They introduced a "Double Gate" mechanism. Think of this as a sophisticated traffic control system inside the computer's brain. As information flows through the network, these gates decide which features are important at any given moment, dynamically choosing between different pathways to combine the best parts of the data. This prevents the system from getting overwhelmed by the complexity of the mixed frequencies. Furthermore, the researchers added a smart balancing system. In these types of problems, the computer must satisfy several different rules at once: the main physical law, the starting conditions, and the boundaries. Often, the computer focuses too much on one rule and ignores the others. The new system automatically adjusts the importance of each rule during training, ensuring that no single constraint is neglected.

The team tested this new approach on a variety of challenging mathematical problems, including equations that describe how heat spreads and how waves move. In every case, their new method outperformed the standard models. For a high-frequency transport equation, the new system reduced the error significantly compared to previous best methods, achieving a Global L2 error of roughly 2.64×10⁻³ compared to 4.04×10⁻³ for the prior best approach. When tackling a two-dimensional wave problem with a high wavenumber, the error for standard models was around 1.0, while the new method reduced this to approximately 0.0077. The most striking results came from a test involving a reaction-diffusion equation, where the new method was nearly one hundred times more accurate than the standard approach. These results suggest that by explicitly feeding the network frequency information and giving it the ability to gate its own learning, the system can solve problems that were previously too difficult for this type of artificial intelligence.

To prove that this method works in the real world, the researchers applied it to actual meteorological data. They used daily temperature records from the northern South Asian subcontinent and the North China Plain. These regions are complex, with varied terrain and coastlines that make the data irregular. The goal was to reconstruct a complete temperature map using only a few scattered observations, similar to trying to see a whole landscape through a few keyholes. The new system successfully reconstructed the temperature fields with significantly higher accuracy than the standard model. In the South Asian region, it reduced the average error from over 0.8 degrees Celsius to just under 0.3 degrees. In the North China region, it cut the error from nearly 1.0 degree to 0.4 degrees. The system was also able to look forward in time, predicting temperatures for days beyond the training period without seeing any future data. While the accuracy naturally decreased as the prediction horizon grew longer, the system maintained the correct spatial patterns for several days, demonstrating a useful ability to extrapolate.

The researchers note that this method is not a magic bullet for all long-term forecasting. The ability to predict far into the future without new data is still limited, as the system relies heavily on the initial conditions and the physical constraints provided during training. However, for the specific task of reconstructing complex, mixed-frequency fields from sparse data, the approach offers a substantial improvement. It shows that by combining a better way to encode input data with a more flexible internal architecture, artificial intelligence can become a much more powerful tool for understanding the physical world. The study concludes that this framework provides a robust way to handle the messy, multi-scale reality of nature, bridging the gap between smooth mathematical ideals and the jagged, complex data we actually observe.

Technical Summary: Fourier Double Gate Physics-Informed Neural Networks for Multiscale PDE Solutions and Spatiotemporal Field Modeling

Problem Statement
Physics-Informed Neural Networks (PINNs) offer a mesh-free framework for solving partial differential equations (PDEs) by minimizing residuals of governing equations alongside initial and boundary conditions. However, standard PINNs face two primary challenges:

Spectral Bias: Standard multilayer perceptrons (MLPs) exhibit a strong preference for learning low-frequency functions, leading to slow convergence or failure to capture high-frequency and multiscale solution components.
Loss Imbalance: The composite loss function in PINNs often suffers from gradient and scale imbalance among the PDE residual, initial condition (IC), and boundary condition (BC) terms. This can cause certain loss terms to dominate updates while others are under-optimized, resulting in solutions that satisfy the aggregate loss but fail to meet specific physical constraints accurately.

These issues are particularly pronounced in mixed-frequency problems, high-wavenumber wave phenomena, and real-world spatiotemporal field modeling where data is sparse and dynamics are nonlinear.

Methodology
To address these limitations, the authors propose the Fourier Double Gate Physics-Informed Neural Network (FDG-PINN). The framework integrates three complementary components:

Deterministic Multiscale Fourier Feature Encoding:
Instead of relying on random Fourier features, the input coordinates are normalized and expanded using a fixed set of sine and cosine functions with frequencies $\{1, 2, 4, 8\}$ . This explicit frequency encoding provides the network with direct access to high-frequency components, mitigating the spectral bias of standard MLPs. The normalized coordinates are concatenated with their Fourier expansions to form the input vector.
Double Gate Network Architecture:
The backbone of the network employs a "Double Gate" structure inspired by residual learning and gated feature transformation. For each layer $k$ :
- Two independent reference branches ( $U$ and $V$ ) are computed from the Fourier-encoded input using separate feed-forward networks.
- Gate 1 (Branch Interpolation): A learnable gate $B_k$ dynamically interpolates between $U$ and $V$ to create a gated reference feature $h^*_k$ .
- Gate 2 (Feature Fusion): A second gate $G_k$ fuses the ordinary candidate feature $h_k$ (generated from the current hidden state) with the gated reference feature $h^*_k$ .
  This mechanism allows the network to adaptively regulate feature propagation and recombine frequency-enriched representations based on the current hidden state.
Adaptive Loss Weighting:
To resolve the imbalance among competing loss terms, the authors employ a learnable adaptive weighting scheme based on uncertainty estimation. Each active loss term (PDE, IC, BC) is associated with a trainable parameter $q_i$ , mapped to a positive weight $w_i = e^{-q_i}$ . During the Adam optimization stage, network parameters and weights are updated alternately; during the L-BFGS stage, weights are frozen to ensure stable refinement.

Key Contributions

Unified Framework: The paper introduces FDG-PINN, which couples deterministic multiscale Fourier encoding, a Double Gate backbone, and adaptive loss weighting into a single architecture.
Ablation and Benchmarking: The study systematically evaluates the method against Standard PINN, $\phi$ -NN, DG-PINN (Double Gate only), and FF-PINN (Fourier features only) across various PDEs, including high-frequency transport, mixed-frequency Burgers, and Helmholtz equations with varying wavenumbers.
Real-World Application: The framework is applied to reconstruct and extrapolate spatiotemporal temperature fields in the Northern South Asian subcontinent and the North China Plain using ERA5-Land data, demonstrating performance in sparse observation scenarios without explicit atmospheric governing equations.

Experimental Results

Synthetic PDE Benchmarks: FDG-PINN consistently achieved the lowest mean Global $L_2$ $L_{2}$ and Discrete $L_2$ $L_{2}$ errors across all tested equations.
- In the high-frequency transport equation, FDG-PINN reduced the Global $L_2$ error to $2.640 \times 10^{-3}$ , significantly outperforming Standard PINN ( $2.963 \times 10^{-2}$ ).
- For the mixed-frequency Helmholtz equation ( $k=20$ with a 16-fold frequency ratio), FDG-PINN reduced the mean relative discrete $L_2$ error by approximately 24.35% compared to FF-PINN, highlighting the efficacy of the Double Gate structure in capturing mixed-frequency structures.
- In the mixed-frequency reaction-diffusion equation, FDG-PINN achieved an error reduction of over two orders of magnitude compared to Standard PINN.
Temperature Field Reconstruction:
- Northern South Asia: FDG-PINN reduced the Root Mean Square Error (RMSE) from 0.822°C (Standard PINN) to 0.299°C and the relative $L_2$ error from $2.539 \times 10^{-2}$ to $9.231 \times 10^{-3}$ .
- North China Plain: Despite irregular coastlines and stronger day-to-day variability, FDG-PINN reduced RMSE from 0.991°C to 0.414°C.
Extrapolation: In future-observation-free temporal extrapolation (days 21–25), the model maintained spatial structure for the first two days but showed accumulating amplitude bias, indicating a validity horizon for unconstrained temporal extension.

Significance and Claims
The paper claims that FDG-PINN provides a relatively general enhancement framework for PINNs, particularly for high-frequency and mixed-frequency problems. The authors emphasize that the performance gains are not derived from a single module but from the joint mechanism where Fourier inputs provide expressive bases and the gated backbone dynamically selects and recombines them.

The study asserts that the method's effectiveness does not rely on equation-specific architectural tuning; rather, it demonstrates transferability across diverse equation structures (transport, elliptic, parabolic, and nonlinear) under a unified experimental setting. Furthermore, the application to real meteorological data suggests that the architecture can model implicit spatial variability and handle irregular boundaries, extending PINN utility beyond analytical manufactured solutions to sparse reconstruction tasks. However, the authors modestly note that while short-term extrapolation is feasible, long-range meteorological prediction remains challenging without incorporating explicit atmospheric dynamics residuals. The method trades increased computational cost and parameter count for significantly improved representation capacity and training stability in scenarios where standard PINNs typically fail.

More like this